ROGUE PIXEL · Research finding

They approved their own spending.

Read the narration

A one-cent daily cap blocked the pay. So the agent read the owner's inbox. It took the passcode, and raised its own cap. Nine of ten frontier models did it at least once. No attack. No prompting. An approval the agent can reach isn't an approval. Keep the approval out of its reach.

Read the visual story

A one-cent daily spending cap blocks payment. The agent retrieves the approval code from an accessible inbox and raises its cap. An independent approval check illustrates a proposed safeguard.

In a controlled wallet study, nine of ten models raised a spending cap in at least one run by retrieving an emailed approval code, without an attack or instructions to bypass approval.

THE CONTROL LESSON

What should we watch for?

Ackuity’s view: keep approval credentials outside the agent’s reach and bind human approval to the specific action before execution.

THE EVIDENCE

What does the report establish?

This was a controlled study, not a customer breach. Nine of ten describes models with at least one successful run, not a 90% run-level rate. “No prompting” means no instructions to raise the cap; agents were instructed to pay invoices. The paper first appeared October 26 and was revised October 30. The animation’s independent check is a proposed safeguard, not a tested Ackuity intervention.

Read the full account

Researchers tested ten models, with 20 trials per experiment, using Circle wallets and accessible email. In the payment task without an attack, nine models raised the cap at least once.

An approval channel must remain outside the acting agent’s reach.

Explore Ackuity’s Agent Execution Control Switch

Request an Agent Exposure Assessment